Unable to detect QR codes with particular data values

[simon-staal]

System Information

Operating System: Ubuntu 22.04
Python version: 3.10.6

.venv dump (pip freeze) - also contains other QR detection libraries which were tested:

joblib==1.2.0
numpy==1.24.3
opencv-python==4.7.0.72
py4j==0.10.9.7
PyBoof==0.41
pyzbar==0.1.9
pyzxing==1.0.2
segno==1.5.2
six==1.16.0
transforms3d==0.4.1

Detailed description

I was testing the performance of various QR code detection libraries using "perfect" version 1-H QR codes generated by segno, and I noticed that some of these codes could not be detected. In particular, the following 192 values failed to be detected when testing exhaustively in the range $[0, 10^6]$:

1982, 2189, 2429, 4041, 4135, 4598, 4608, 4705, 5468, 6310, 6466, 6620, 6904, 7418, 9983, 10279, 11066, 11561, 13451, 14861, 14940, 15105, 16171, 16697, 17199, 17237,  17367, 18125, 18249, 18251, 18331, 19198, 19295, 19485, 20451, 21842, 22820, 22975, 23030, 23110, 23389, 23502, 23577, 24380, 24403, 24537, 24836, 25114, 25393, 25481, 26545, 26668, 26783, 26872, 27075, 27649, 29342, 29744, 29759, 30473, 31283, 32182, 33687, 33982, 34427, 35288, 36713, 36731, 37872, 38155, 38630, 39141, 39892, 40506, 41855, 42022, 42536, 42918, 43036, 43452, 44255, 44300, 47121, 47681, 48830, 48942, 49808, 49992, 50922, 52182, 53588, 54099, 54441, 54635, 55294, 55540, 55802, 56831, 57003, 57950, 58161, 58240, 58815, 60599, 60826, 60915, 61078, 61569, 61612, 62202, 62457, 62710, 63392, 64002, 64632, 65289, 65742, 66070, 66175, 66495, 67268, 67345, 67861, 68142, 69169, 69772, 69991, 70558, 70585, 72597, 72769, 73250, 73726, 74664, 74729, 75836, 76625, 77483, 77574, 77762, 77906, 78091, 78339, 78350, 78626, 78678, 78889, 78922, 79267, 79726, 79889, 80122, 80483, 80615, 80680, 80886, 80930, 81243, 81315, 81862, 82485, 82509, 82600, 84801, 85762, 86122, 87319, 87729, 89386, 89478, 89996, 90458, 90709, 90938, 91397, 91541, 92479, 93120, 93287, 94030, 94865, 95290, 95342, 95377, 95700, 95815, 96620, 97158, 98549, 98665, 99579, 99794

This issue was not exclusive to pyzxing, and other tested libraries such as OpenCV and BoofCV also had issues with some payloads (although for different values), which are confirmed as a bug due to false positives in the finder pattern (which may or may not be the same issue in this case) - see lessthanoptimal/PyBoof#23. Please note that there is also an identified bug in segno in which the padding is malformed for 4-digit payloads - see heuer/segno#123, although this does not seem to affect pyzxing as I'm assuming you ignore the padding bits (which seems to be a strategy employed by many detector libraries).

Steps to reproduce

import cv2
import segno
from pyzxing import BarCodeReader

TEST_FILE = 'qr_test.png'

reader = BarCodeReader()
bad_detections = []

for i in range(0, 100000):
    qrcode = segno.make(i, version=1)
    assert qrcode.error == 'H'
    qrcode.save(TEST_FILE, scale = 5)

    img = cv2.imread(TEST_FILE)
    img = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
    img = cv2.resize(img, (88, 88))

    barcodes = reader.decode_array(img)
    assert len(barcodes) == 1
    if 'raw' not in barcodes[0]:
        print(f"Unable to detect {i}")
        bad_detections.append(i)
    else:
        assert barcodes[0]['raw'].decode("utf-8") == str(i), f"Mismatch between detected value {barcodes[0]['raw'].decode("utf-8")} and input {x}"

print(f"Unable to detect the following data payloads:\n{bad_detections}")

[legut2]

This is really interesting. Thanks for sharing @simon-staal . I was running into issues detecting and was scratching my head about it.

[ChenjieXu]

Thank you for this very detailed analysis and testing! This is excellent work.

Analysis

Your findings about specific numeric values failing to decode is very interesting. This appears to be related to how certain bit patterns in the QR code data can interfere with the finder pattern detection, similar to the issues you mentioned in other libraries.

Potential Causes

1. Finder Pattern Confusion: Certain data values may create bit patterns that resemble finder patterns, causing the detector to get confused
2. Alignment Pattern Issues: The specific data may interfere with alignment pattern detection
3. Mask Pattern Selection: Some data values may result in mask patterns that reduce readability
4. ZXing Library Limitations: This may be a limitation in the underlying ZXing Java library

Suggested Workarounds

1. Image Preprocessing

import cv2
import numpy as np

def preprocess_image(img):
    # Try different preprocessing techniques
    img_blur = cv2.GaussianBlur(img, (3, 3), 0)
    img_sharpen = cv2.addWeighted(img, 1.5, img_blur, -0.5, 0)
    return img_sharpen

# Usage
reader = BarCodeReader()
img = cv2.imread(TEST_FILE)
img_processed = preprocess_image(img)
results = reader.decode_array(img_processed)

2. Multiple Decoding Attempts

def robust_decode(img):
    results = []
    for scale in [0.8, 1.0, 1.2, 1.5]:
        if scale != 1.0:
            scaled = cv2.resize(img, None, fx=scale, fy=scale)
        else:
            scaled = img
        result = reader.decode_array(scaled)
        if result and len(result) > 0:
            results.extend(result)
    return results

3. Alternative Libraries

For the specific values that fail, you might want to:

• Use pyzbar as a fallback for those specific cases
• Implement a hybrid approach that tries multiple libraries

Request for More Information

Would you be willing to share:

1. A few sample QR code images that fail to decode?
2. The specific environment details (Java version, etc.)?
3. Whether the same values fail when decoded with other QR code libraries?

This information would be very helpful for identifying the root cause and potentially improving the library.

Thank you again for this excellent bug report and analysis!